Who can tell me answer?
", "messages": [ { "id": "437653", "postDate": "12/12/2018 09:26:54", "content": "Who can tell me answer?
", "rawMarkdown": "Who can tell me answer?", "votes": null }, { "id": "437674", "postDate": "12/12/2018 09:52:42", "content": "It looks like Gradient Boosting Decision Tree often does its job better than NN does for tabular data, but NN predictions may be useful as a part of ensemble. I think nobody would try to use LightGBM instead of NN, for example, to recognize images.
", "rawMarkdown": "It looks like Gradient Boosting Decision Tree often does its job better than NN does for tabular data, but NN predictions may be useful as a part of ensemble. I think nobody would try to use LightGBM instead of NN, for example, to recognize images.", "votes": null }, { "id": "437739", "postDate": "12/12/2018 12:14:15", "content": "did you try it ? If yes, Was it yield good score?
", "rawMarkdown": "did you try it ? If yes, Was it yield good score?", "votes": null }, { "id": "437746", "postDate": "12/12/2018 12:36:49", "content": "I'm a kaggle novice, and thank you for answering my questions. I'm going to try to compare neural networks and GBDT on this task and see what will happen
", "rawMarkdown": "I'm a kaggle novice, and thank you for answering my questions. I'm going to try to compare neural networks and GBDT on this task and see what will happen", "votes": null }, { "id": "437749", "postDate": "12/12/2018 12:38:05", "content": "no , I'm going to do this.
", "rawMarkdown": "no , I'm going to do this.", "votes": null }, { "id": "437754", "postDate": "12/12/2018 12:43:52", "content": "Hope you will share it :D
", "rawMarkdown": "Hope you will share it :D", "votes": null }, { "id": "437859", "postDate": "12/12/2018 16:17:48", "content": "I've extensively tried NN for this task. They don't work well because as I mentioned in my \"Some Concerns\" post here: https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73425, the data provided has alsmost 0 relationship with the target variables and the resulting size of the error even for the top of the leaderboard is huge. NN are too powerful or \"greedy\" and can't converge given the near total lack of signal in the provided data.
", "rawMarkdown": "I've extensively tried NN for this task. They don't work well because as I mentioned in my \"Some Concerns\" post here: https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73425, the data provided has alsmost 0 relationship with the target variables and the resulting size of the error even for the top of the leaderboard is huge. NN are too powerful or \"greedy\" and can't converge given the near total lack of signal in the provided data.", "votes": null }, { "id": "438022", "postDate": "12/13/2018 01:20:18", "content": "Very helpful answer. Nice work. Thank you for answering my question
", "rawMarkdown": "Very helpful answer. Nice work. Thank you for answering my question", "votes": null }, { "id": "438025", "postDate": "12/13/2018 01:28:56", "content": "but the url you given can't open, there is a 404
", "rawMarkdown": "but the url you given can't open, there is a 404", "votes": null }, { "id": "438423", "postDate": "12/13/2018 16:25:57", "content": "May be DAE would help (refer to the famous Michael Jahrer's solution on Porto Seguro competition)
", "rawMarkdown": "May be DAE would help (refer to the famous Michael Jahrer's solution on Porto Seguro competition)", "votes": null }, { "id": "438622", "postDate": "12/14/2018 00:25:52", "content": "It's called \"Some Concerns\"
\n\nhttps://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73425
", "rawMarkdown": "It's called \"Some Concerns\"\n\nhttps://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73425", "votes": null }, { "id": "438834", "postDate": "12/14/2018 08:59:12", "content": "I tried using Deep Neural Nets in the beginning and failed miserably. Even with all the effort to avoid overfitting with regularization, drop out, and all that jazz nothing really helped.
", "rawMarkdown": "I tried using Deep Neural Nets in the beginning and failed miserably. Even with all the effort to avoid overfitting with regularization, drop out, and all that jazz nothing really helped.", "votes": null }, { "id": "439850", "postDate": "12/16/2018 14:21:37", "content": "Yea me too. It has to do with what whamp said: \"more than HALF of the RMSE is due to about\n2000 outliers with loyalty scores of ~ -33\"
\n\nAnd predicting these outliers is crucial!
\n\nSo models that have grided outputs will perform better. NN's are much less capable of that compared to tree-based models.
", "rawMarkdown": "Yea me too. It has to do with what **whamp** said: \"more than HALF of the RMSE is due to about\n2000 outliers with loyalty scores of ~ -33\" \n\nAnd predicting these outliers is crucial!\n\nSo models that have grided outputs will perform better. NN's are much less capable of that compared to tree-based models.", "votes": null }, { "id": "440767", "postDate": "12/18/2018 01:02:00", "content": "Makes sense!
", "rawMarkdown": "Makes sense!", "votes": null }, { "id": "441245", "postDate": "12/18/2018 13:20:23", "content": "thank you very much
", "rawMarkdown": "thank you very much", "votes": null } ], "comments": [ { "id": 437674, "author_name": "lumaxy", "author_url": "", "post_date": "12/12/2018 09:52:42", "content": "It looks like Gradient Boosting Decision Tree often does its job better than NN does for tabular data, but NN predictions may be useful as a part of ensemble. I think nobody would try to use LightGBM instead of NN, for example, to recognize images.
", "votes": null, "replies": [ { "id": 437746, "author_name": "lgcxjtu", "author_url": "", "post_date": "12/12/2018 12:36:49", "content": "I'm a kaggle novice, and thank you for answering my questions. I'm going to try to compare neural networks and GBDT on this task and see what will happen
", "votes": null, "replies": [] } ] }, { "id": 437739, "author_name": "qnkhuat", "author_url": "", "post_date": "12/12/2018 12:14:15", "content": "did you try it ? If yes, Was it yield good score?
", "votes": null, "replies": [ { "id": 437749, "author_name": "lgcxjtu", "author_url": "", "post_date": "12/12/2018 12:38:05", "content": "no , I'm going to do this.
", "votes": null, "replies": [] }, { "id": 437754, "author_name": "qnkhuat", "author_url": "", "post_date": "12/12/2018 12:43:52", "content": "Hope you will share it :D
", "votes": null, "replies": [] } ] }, { "id": 437859, "author_name": "hampsonw", "author_url": "", "post_date": "12/12/2018 16:17:48", "content": "I've extensively tried NN for this task. They don't work well because as I mentioned in my \"Some Concerns\" post here: https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73425, the data provided has alsmost 0 relationship with the target variables and the resulting size of the error even for the top of the leaderboard is huge. NN are too powerful or \"greedy\" and can't converge given the near total lack of signal in the provided data.
", "votes": null, "replies": [ { "id": 438022, "author_name": "lgcxjtu", "author_url": "", "post_date": "12/13/2018 01:20:18", "content": "Very helpful answer. Nice work. Thank you for answering my question
", "votes": null, "replies": [] }, { "id": 438025, "author_name": "lgcxjtu", "author_url": "", "post_date": "12/13/2018 01:28:56", "content": "but the url you given can't open, there is a 404
", "votes": null, "replies": [] }, { "id": 438423, "author_name": "serigne", "author_url": "", "post_date": "12/13/2018 16:25:57", "content": "May be DAE would help (refer to the famous Michael Jahrer's solution on Porto Seguro competition)
", "votes": null, "replies": [] }, { "id": 438622, "author_name": "hampsonw", "author_url": "", "post_date": "12/14/2018 00:25:52", "content": "It's called \"Some Concerns\"
\n\nhttps://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73425
", "votes": null, "replies": [] }, { "id": 441245, "author_name": "lgcxjtu", "author_url": "", "post_date": "12/18/2018 13:20:23", "content": "thank you very much
", "votes": null, "replies": [] } ] }, { "id": 438834, "author_name": "enzoamp", "author_url": "", "post_date": "12/14/2018 08:59:12", "content": "I tried using Deep Neural Nets in the beginning and failed miserably. Even with all the effort to avoid overfitting with regularization, drop out, and all that jazz nothing really helped.
", "votes": null, "replies": [ { "id": 439850, "author_name": "antonisst", "author_url": "", "post_date": "12/16/2018 14:21:37", "content": "Yea me too. It has to do with what whamp said: \"more than HALF of the RMSE is due to about\n2000 outliers with loyalty scores of ~ -33\"
\n\nAnd predicting these outliers is crucial!
\n\nSo models that have grided outputs will perform better. NN's are much less capable of that compared to tree-based models.
", "votes": null, "replies": [] }, { "id": 440767, "author_name": "enzoamp", "author_url": "", "post_date": "12/18/2018 01:02:00", "content": "Makes sense!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "437653": "Who can tell me answer?", "437674": "It looks like Gradient Boosting Decision Tree often does its job better than NN does for tabular data, but NN predictions may be useful as a part of ensemble. I think nobody would try to use LightGBM instead of NN, for example, to recognize images.", "437739": "did you try it ? If yes, Was it yield good score?", "437746": "I'm a kaggle novice, and thank you for answering my questions. I'm going to try to compare neural networks and GBDT on this task and see what will happen", "437749": "no , I'm going to do this.", "437754": "Hope you will share it :D", "437859": "I've extensively tried NN for this task. They don't work well because as I mentioned in my \"Some Concerns\" post here: https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73425, the data provided has alsmost 0 relationship with the target variables and the resulting size of the error even for the top of the leaderboard is huge. NN are too powerful or \"greedy\" and can't converge given the near total lack of signal in the provided data.", "438022": "Very helpful answer. Nice work. Thank you for answering my question", "438025": "but the url you given can't open, there is a 404", "438423": "May be DAE would help (refer to the famous Michael Jahrer's solution on Porto Seguro competition)", "438622": "It's called \"Some Concerns\"\n\nhttps://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73425", "438834": "I tried using Deep Neural Nets in the beginning and failed miserably. Even with all the effort to avoid overfitting with regularization, drop out, and all that jazz nothing really helped.", "439850": "Yea me too. It has to do with what **whamp** said: \"more than HALF of the RMSE is due to about\n2000 outliers with loyalty scores of ~ -33\" \n\nAnd predicting these outliers is crucial!\n\nSo models that have grided outputs will perform better. NN's are much less capable of that compared to tree-based models.", "440767": "Makes sense!", "441245": "thank you very much" }, "source": "meta" }